510(k) and De Novo

A 510(k) is a premarket submission showing the FDA that a new device is substantially equivalent to a legally marketed device, called the predicate 1. A De Novo request is the route for a novel low- to moderate-risk device with no suitable predicate: if the FDA grants it, the agency creates a new device classification 2.

Why do these pathways matter in healthcare?

They are how almost every AI device reaches clinicians. Of 168 machine-learning-enabled Class II devices the FDA authorized in 2024, 94.6% came through 510(k) and 5.4% through De Novo. Median review time was 151 days for 510(k) and 372 days for De Novo 3. For premarket approval and the wider regulatory category, see Software as a Medical Device.

The pathway shapes the evidence. A 510(k) answers a comparative question: is this device as safe and effective as one already on the market? In a 2021 review of 130 FDA-authorized AI devices, 126 had been evaluated only with retrospective studies at submission 5.

How does each pathway work?

510(k). The sponsor names a predicate and shows the new device has the same intended use, and either the same technological characteristics or different ones that raise no different questions of safety and effectiveness 1. A cleared device can then serve as a predicate for the next one, so devices form chains.

De Novo. The sponsor shows that general controls, or general and special controls, give reasonable assurance of safety and effectiveness. If the FDA grants the request, a new classification regulation is created, and the device may serve as a predicate for future 510(k)s of the same type 2.

The first autonomous AI diagnostic system, IDx-DR, reached the market through De Novo in April 2018. It reads retinal images for more than mild diabetic retinopathy without a clinician interpreting the image 6. The trial behind it enrolled 900 people with diabetes at primary care sites and reported sensitivity of 87.2% and specificity of 90.7% 7.

Where does this show up today?

Predicate chains can cross the line between conventional software and AI. An audit of AI/ML devices cleared from 2019 to 2021 found that more than a third traced back to non-AI/ML devices in the first generation of their predicate networks 4. In the 2024 cohort, 64.5% of predicates cited by ML-enabled 510(k) devices were themselves ML-enabled 3.

A recent example: in December 2025 the FDA cleared UpDoc, an insulin-management device with a patient-facing large language model interface, through 510(k) 8. Its predicate was an insulin-dose calculator cleared in 2019 8, a story told in the first FDA-cleared patient-facing LLM.

Common misunderstandings

Cleared means approved. Devices that pass 510(k) are cleared, De Novo requests are granted, and approval refers to premarket approval. The words track different standards.

Equivalence means the AI was compared with other AI. The predicate may contain no AI at all 3 4.

Clearance covers future versions. Changes to a cleared model can require a new submission. An authorized predetermined change control plan lets pre-specified modifications ship without one 9.

What to ask a vendor

• What is the 510(k) or De Novo number? Read the decision summary with how to read an FDA clearance summary.
• What was the predicate, and did it use machine learning?
• What evidence supported the submission: retrospective or prospective, how many sites, which populations?
• Is the version we would deploy the cleared version, and does it carry an authorized PCCP?
• Does the cleared indication match how we plan to use it? A model card should state the intended use.

Related terms

• Software as a Medical Device: the category these pathways regulate.
• Predetermined change control plan: how a cleared AI device updates.
• FDA-cleared AI devices by year and specialty: the running counts.